{"id":"e73ea5f0-6d3e-44a0-846a-e49936846a5d","arxiv_id":"2603.05774","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A single-loop softmax-weighted switching-gradient method solves constrained federated minimax problems at Õ(ε^{-4}) oracle complexity with high-probability guarantees and partial-participation analysis.","lead":"The paper introduces a primal-only Softmax-Weighted Switching Gradient algorithm for federated learning that optimizes the worst-case client loss while enforcing stochastic constraints, without dual variables. It gives standard stochastic rates under full participation and a high-probability analysis for partial participation under a client-gap assumption.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Convexity (Ass. 4.1) is indispensable to every main theorem and the Bregman arguments, yet the primary non-convex DNN experiment is presented as empirical support.","rationale":"The reader correctly isolates Assumption 4.1 as the load-bearing premise for the three main theorems and notes the theory–experiment mismatch. The partial-participation Assumption 4.9 is secondary; the convexity gap already prevents the non-convex experiments from validating the stated guarantees. No deeper algebraic flaw in the high-probability arguments was found, so the CONDITIONAL verdict and moderate confidence remain appropriate. The abstract’s unsubstantiated RL claim is a presentation defect but does not alter the mathematical status of the theorems.","tokens_in":62411,"tokens_out":500,"duration_ms":19630,"concrete_test":"Replace the DNN in the Adult fair-classification experiment (Section 5, Table 2) by a convex logistic model with the same demographic-parity constraint; re-run the three methods under identical participation and local-update settings. If the qualitative ranking or the constraint-satisfaction curves change materially, the non-convex results cannot be cited as support for Theorems 4.6–4.10.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Main Theorems 4.6, 4.7 and 4.10 (and the entire error decomposition in Appendices C–E) rest on Assumption 4.1 that every local f_i and g_i is convex. Convexity is used to guarantee non-negative Bregman divergences, to invoke the three-point identity (Lemma B.1), and to pass from the averaged feasible iterates to the claimed optimality-gap and feasibility bounds via Jensen. Section 5’s fair-classification experiment, however, employs a deep network whose landscape is explicitly called “highly non-convex and non-smooth.” Consequently the optimality-gap and feasibility guarantees do not apply to the very setting used to “verify the efficacy” of the algorithm. The NP experiment is convex, but the paper’s strongest empirical claim (competitive performance against primal-dual/penalty baselines on a modern non-convex task) lies outside the theorem statements. Without either a non-convex analysis or a clear disclaimer, the experimental section cannot be read as corroboration of the central complexity claims.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes a Softmax-Weighted Switching Gradient method for distributed stochastic minimax optimization with stochastic constraints in federated learning. It replaces the hard client-wise max with a temperature-controlled softmax, then switches between objective and constraint gradients according to a feasibility check, without dual variables. Under convexity, Lipschitz continuity, diameter, and sub-Gaussian assumptions, Main Theorems 4.6 and 4.7 give high-probability O(ε) optimality-gap and feasibility guarantees for full participation at the standard Õ(ε^{-4}) oracle complexity, with an improved O(log 1/δ) dependence. Under an additional stochastic-superiority assumption (4.9), Theorem 4.10 extends the analysis to partial participation. The softmax temperature lower bound is tightened by dropping a global loss bound. Experiments cover Neyman-Pearson classification (convex) and fair classification with a DNN (non-convex).","tokens_in":62711,"tokens_out":1278,"duration_ms":12879,"significance":"If the analysis holds, the work supplies a clean primal-only single-loop alternative to primal-dual and penalty methods for constrained worst-case federated learning, with an explicit three-way error decomposition (optimization / estimation / sampling) and a sharper high-probability rate. The relaxation of the bounded-loss assumption for the softmax parameter and the partial-participation extension under a stochastic-dominance condition are genuine technical contributions relative to prior softmax-minimax and switching-subgradient literature. The appendices contain a full high-probability derivation under the stated assumptions, which is a clear strength.","major_comments":[{"comment":"Assumption 4.1 (convexity of every local f_i and g_i) is load-bearing for Main Theorems 4.6, 4.7 and 4.10: non-negative Bregman divergences, the three-point identity (Lemma B.1), and Jensen passage from averaged feasible iterates to the claimed gap/feasibility bounds all require it. Section 5’s fair-classification experiment, however, uses a deep network whose landscape is described as “highly non-convex and non-smooth.” The optimality-gap and feasibility guarantees therefore do not apply to the setting used to claim competitive performance against primal-dual/penalty baselines. Either a non-convex analysis (or stationarity guarantee) or an explicit disclaimer that the DNN results are purely empirical is needed; without it the experimental section cannot be read as corroboration of the central complexity claims.","section":null},{"comment":"For partial participation, Assumption 4.9 (relative gaps F−f_i and G−g_i stochastically dominated by Unif[0,σ]) is the structural premise that makes client-sampling error O(σ / |ln(1−r)| n) controllable (Theorem 4.10 and the error-decomposition box). The assumption is non-standard, hard to verify on real federated data, and not checked or even estimated in the experiments (m/n = 0.5). The paper should either provide diagnostic evidence that the ordered-gap condition holds approximately on the Adult/Breast-Cancer partitions, or clearly mark the partial-participation rate as conditional on this premise and discuss failure modes when it is violated.","section":null},{"comment":"Abstract and introduction list “federated safe reinforcement learning” among the experimental tasks, yet Section 5 and Appendix F only report NP classification and fair classification. Either the RL experiment must be restored with the same rigor as the other two, or the claim must be removed from the abstract and contribution list so that the empirical scope matches the manuscript body.","section":null}],"minor_comments":[{"comment":"Abstract vs. body inconsistency on experimental tasks (safe RL mentioned only in abstract) should be resolved for consistency.","section":null},{"comment":"Figure 2 caption and surrounding text discuss α-sensitivity; the practical switch threshold is relaxed from ε/2 to ε/1.1 in Appendix F without a corresponding statement in the main algorithm description—clarify that the theory uses ε/2 while practice uses a milder constant.","section":null},{"comment":"Notation for masked softmax (Eq. 10) and the localized maxima F(w; I_k), G(w; I_k) (Eq. 11) is introduced late; a short pointer in Section 2 would help.","section":null},{"comment":"Typos: “ganrantees” (p. 5), “fo the worst-case” (p. 5), “condtional” / “defintion” in the appendices; also “V on Neumann” in Related Work.","section":null},{"comment":"Related Work (Appendix G) is thorough but could more sharply position the O(log 1/δ) improvement against Lan & Zhou (2020b) and the α bound against Wang et al. (2023) in a short comparison table or paragraph in the main text.","section":null}],"recommendation":"major_revision","confidential_remarks":"The convexity–experiment mismatch is the main reason I recommend major rather than minor revision; the theory itself looks carefully executed under its assumptions. If the authors add a clear non-convex disclaimer (or a weak-convexity extension) and either justify or demote Assumption 4.9, the paper would be a solid contribution for a methods-oriented ML venue. The missing RL experiment is a smaller but easy-to-fix presentation issue."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The real contribution is a single-loop primal-only method that combines softmax weighting of the client max with a Polyak-style switch on the constraint, plus a full high-probability analysis that gets the standard Õ(ε^{-4}) stochastic rate and a cleaner O(log 1/δ) factor. They also drop the global loss bound that Wang et al. needed for the softmax temperature and get a tighter α ≳ ln n / ε' (or ln m under partial participation). That package is useful for people who actually have to keep dual variables from drifting under client sampling.\n\nThe math is carefully written. Appendices C–E spell out the Bregman three-point identities, sub-Gaussian concentration, and supermartingale arguments under the stated convexity/Lipschitz/diameter/sub-Gaussian assumptions. The error split into optimization / estimation / sampling terms is explicit, and the partial-participation extension via the stochastic-superiority (FSD) client-gap assumption is a genuine (if non-standard) technical device. The NP classification experiment sits inside the theory and looks clean against penalty and primal-dual baselines.\n\nThe soft spots are real but not fatal. Every main theorem leans on convexity of every local f_i and g_i; the fair-classification experiment is a deep net the authors themselves call highly non-convex and non-smooth, so those plots cannot be read as corroboration of the complexity claims. The abstract also promises federated safe RL that never appears in Section 5. Assumption 4.9 is hard to check on real data, and there is no released code. None of that breaks the convex full-participation theory; it just means the empirical section over-reaches.\n\nThis is for the constrained-FL / Group-DRO / switching-subgradient crowd. A serious referee should see it. I would cite the dual-free rate and the α bound if I were writing in that area, and I would bring the partial-participation analysis to reading group if we are talking about client sampling noise. Send it to review; ask them to either add a non-convex disclaimer or a weak-convexity extension and to drop the missing RL claim.","headline":"Solid dual-free constrained federated minimax with clean high-prob rates; theory is convex, main DNN experiment is not.","tokens_in":63379,"tokens_out":550,"would_cite":true,"duration_ms":8058,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A single-loop primal-only switching method with softmax weights solves federated worst-case minimax problems under stochastic constraints at the standard Õ(ε⁻⁴) rate.","keywords":["federated learning","stochastic minimax optimization","stochastic constraints","switching gradient method","softmax smoothing","partial participation","high-probability convergence","primal-only method"],"falsifier":"On a convex federated Neyman-Pearson or fair-classification instance with known optimum, run the algorithm with the prescribed step-sizes and batch sizes and check whether the averaged feasible iterate simultaneously meets both ε-optimality and ε-feasibility at the predicted sample complexity; failure under verified convexity and sub-Gaussian noise would refute the claim.","tokens_in":63247,"feed_emoji":"⚖️","tokens_out":946,"duration_ms":10101,"temperature":0.7,"pith_summary":"Federated models that only minimize average loss can leave some clients with poor performance, and adding client-specific safety or fairness constraints usually forces dual variables that drift and oscillate under partial participation. This paper shows that a simple first-order switching rule, smoothed by temperature-controlled softmax weights over clients, can optimize the worst-case client loss while keeping the worst-case constraint feasible, without any dual variables or nested loops. Under standard convexity and sub-Gaussian noise assumptions the method attains a unified ε-accuracy on both optimality gap and feasibility at the classical stochastic complexity Õ(ε⁻⁴), with a sharper high-probability dependence O(log 1/δ). Partial participation is handled by an extra stochastic-superiority assumption that bounds client sampling noise; the same analysis also yields a tighter lower bound on the softmax temperature once boundedness of the losses is dropped. Experiments on Neyman-Pearson classification and fair classification illustrate that the primal-only scheme is less sensitive to hyper-parameters than penalty or primal-dual baselines.","feed_headline":"Primal-only switching hits federated worst-case rates without duals","feed_subtitle":"Softmax weights plus a feasibility switch give Õ(ε⁻⁴) on both gap and constraint under partial participation","key_machinery":"Softmax-weighted switching: at each round the server forms temperature-α softmax weights over participating clients for both objective and constraint values; if the softmax-mean constraint is below a tightened tolerance it steps on the weighted objective gradients, otherwise on the weighted constraint gradients, then averages only the feasible iterates.","core_discovery":"A single-loop Softmax-Weighted Switching Gradient algorithm, which alternates between a softmax-smoothed worst-case objective gradient and a softmax-smoothed constraint gradient according to a feasibility trigger, returns an averaged iterate that simultaneously satisfies optimality gap ≤ ε and feasibility violation ≤ ε (up to a vanishing sampling term under partial participation) with probability 1-δ at the standard Õ(ε⁻⁴) stochastic oracle complexity.","pith_inferences":["The same switching-plus-softmax idea may extend, with weaker rates, to the weakly convex regime already studied for unconstrained switching methods, matching the non-convex fair-classification experiments.","If the stochastic-superiority gap σ can be estimated on-line from client statistics, one could adaptively choose the participation ratio to keep sampling error below a target.","The unified error decomposition (optimization + estimation + sampling) supplies a practical checklist for allocating communication rounds, local batch sizes and client sampling rates."],"forward_implications":["Primal-only switching can replace dual-variable methods in federated settings that require both worst-case performance and client-wise safety or fairness constraints.","The high-probability rate improves from O(log^{2} 1/δ) to O(log 1/δ), giving tighter sample sizes for the same failure probability.","Dropping the usual bounded-loss assumption yields a strictly smaller admissible range for the softmax temperature α.","Under the stochastic-superiority condition the same rates extend to partial participation, recovering the full-participation bound as the participation ratio tends to one."],"fun_headline_variants":["Softmax switch hits dual-free federated worst-case rates","Single-loop primal switch reaches Õ(ε⁻⁴) on gap and constraints","Feasibility-triggered softmax gradients stabilize constrained FL","Softmax-weighted switch bypasses duals for federated minimax","Primal switching with softmax weights unifies gap and feasibility"],"cache_read_input_tokens":49280,"weakest_assumption_plain":"Every local objective and constraint function must be convex; without that the optimality and feasibility guarantees do not hold.","fun_headline_variants_meta":{"raw":{"variants":["Softmax switch hits dual-free federated worst-case rates","Single-loop primal switch reaches Õ(ε⁻⁴) on gap and constraints","Feasibility-triggered softmax gradients stabilize constrained FL","Softmax-weighted switch bypasses duals for federated minimax","Primal switching with softmax weights unifies gap and feasibility"]},"model":"grok-4.5","effort":"low","cost_usd":0.004672,"raw_usage":{"total_tokens":1332,"prompt_tokens":782,"num_sources_used":0,"completion_tokens":73,"cost_in_usd_ticks":46720000,"prompt_tokens_details":{"text_tokens":782,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":477,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":782,"tokens_out":73,"duration_ms":4495,"temperature":1.0,"reasoning_tokens":477,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-15T14:16:13.869069+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On a convex federated Neyman-Pearson or fair-classification instance with known optimum, run the algorithm with the prescribed step-sizes and batch sizes and check whether the averaged feasible iterate simultaneously meets both ε-optimality and ε-feasibility at the predicted sample complexity; failure under verified convexity and sub-Gaussian noise would refute the claim.","supporting_citations":[],"review_version":1}